Your prompt is the ceiling on every answer you get
Most prompt tools wrap your words in the same handwritten template and call it a day. This one reads your task, picks a method out of 6,235 taken from published AI research, and writes against the paper behind it. No account, no email, no card.
Why this number is low
It is raw similarity from one retrieval pass. No specialist read the paper behind this technique. No judge compared it against alternatives. It looks related to your wording; nobody verified it works for your problem.
- 1 retrieval pass, nearest match wins
- 0 papers read in full
- 1 model, the smallest we run
- No alternatives, no comparison
- Gone when you close the tab
- 10 specialists read the full paper behind each candidate
- A reasoning judge ranks your Top-3 with evidence
- You pick, with strengths and weaknesses shown
- 7 models, up to GPT-5.6 Sol and Claude Opus 4.8
- Saved to history, regenerable, comparable side by side
How this works
Most prompt generators wrap your text in a template someone wrote by hand: add a role, add "think step by step", ask for a structure, return it. That is fine, and it is also the same output for every task.
This one starts somewhere else. Behind it is a corpus of 6,235 prompting techniques extracted from published AI research, each one carrying what it does, when it applies, and what effect the authors measured. When you describe a task, the tool searches that corpus, takes the closest technique, and writes a prompt that actually applies it to your specific case.
So the output is not a template. It is one particular method, chosen because it matched what you asked for, with a link to the paper so you can check it yourself.
Why the confidence number is low, and why we show it
A single similarity search is a weak instrument. It finds techniques that look related to your wording, which is not the same as techniques that work for your problem. Two techniques can sit at nearly the same distance from your query and be worlds apart in usefulness, and no retrieval score can tell them apart.
That is what the ten specialist agents and the judge exist for in the paid product: each candidate gets read in full, then compared. We show you the raw number here because the honest version of a free tool is one that tells you where its own limits are.
What a prompting technique actually is
A technique is a repeatable structure for a request, not a magic phrase. A few examples of the families in the corpus:
- Few-shot prompting. Show two or three worked examples of the input-output pair you want before asking for the real one. Strong for format-sensitive and extraction work.
- Chain of thought. Ask for the reasoning before the answer. Well known, and worth knowing that the payoff has shrunk sharply on 2025-2026 reasoning models, which already reason internally. On some of them it measurably hurts.
- Self-consistency. Generate several independent answers and take the one that recurs. Expensive, and the standard fix when the same prompt keeps giving different results.
- Decomposition. Break one large request into ordered sub-tasks, each verifiable. The usual answer when a model handles a long task confidently and wrongly.
- Critic and reviewer. Have the model produce, then criticise its own output against explicit criteria, then revise. Effective against confident filler.
The corpus holds roughly a thousand distinct patterns beyond these, most of them narrow and specific, which is exactly why picking by hand does not scale.
When a technique will not save you
Worth saying plainly, because most pages in this category will not: prompting cannot fix a task that is underspecified, and it cannot supply context the model does not have. If the model does not know your product, your data or your constraints, no structure around the request will invent them correctly. The technique layer matters after the context layer is right, not instead of it.
Equally, on current reasoning models several once-standard tricks now add little. Persona instructions, politeness, and offers of a reward have been measured repeatedly and come out close to noise. A tool that sells you those is selling 2023.
Where the techniques come from
The corpus is assembled from published research, mostly arXiv preprints. Every one of the 6,235 entries carries a paper identifier, which is why the tool can link you to the source of whatever it just used. Roughly 95% of the corpus rests on work published in 2025 and 2026, and more than 2,300 of the techniques carry a measured result rather than a qualitative claim.
It grows continuously as new work is published, which is the part that does not fit in a static prompt pack.
Use it from your editor
The same matcher is exposed over MCP, so a client that speaks it (Claude Desktop, Claude Code, Cursor and others) can call it directly. Ask your assistant for a prompting technique and it fetches one, with the paper behind it, without you leaving the editor. Free, no account, same per-IP limit as this page.
claude mcp add --transport http getdecision https://api.getdecision.ai/mcp/
Any MCP client works; the command above is the Claude Code one. The tool is
pick_prompting_technique, it takes the task in plain language, and it returns the
technique, a ready prompt, the arXiv link and the confidence of the match so your assistant can
tell a weak result from a good one.
Frequently asked
Is it really free, and do I need an account?
Yes and no respectively. No signup, no email, no card. There is a per-minute rate limit so the page survives being crawled and scripted.
Which model does the free tool use?
The smallest and cheapest in our lineup, Gemini 2.5 Flash-Lite. That is deliberate: it keeps the page free. The full product lets you generate on seven models including GPT-5.6 Sol and Claude Opus 4.8.
Can I use the generated prompt anywhere?
Yes. The prompts are yours, with any model, commercial use included.
Why does it take about ten seconds?
Because it loads the actual source document behind the chosen technique and writes against it, rather than filling a template. A template would be instant and would be the same for everyone.
Does it work in languages other than English?
Yes. Describe the task in your own language and the prompt will follow it.
Can a paid account use this from an editor too?
Not yet. The MCP server exposes the free single-pass matcher only. Running the full ten-specialist analysis from inside a client, against your own credit balance, is a separate piece of work and it is not built.
How is this different from asking ChatGPT to write my prompt?
A model writing its own prompt draws on its general sense of what a good prompt looks like. This draws on a specific, cited method selected for your task, and tells you which one it used so you can disagree.
Related
- What GetDecision does, the full ten-specialist pipeline
- Create a free account and run one full search